AI & Computer Science · Weekly AI Papers
Yann LeCun & Randall Balestriero: The Platonic Representation Hypothesis with JEPAs
LISTEN TO THIS EPISODE
Can models trained separately on images and language arrive at the same representation of the world?
Randall Balestriero and Yann LeCun derive alignment up to a rotation for an idealized JEPA setting with shared positive pairs and additional assumptions.
They distinguish good views of the same scene from views that preserve the peculiarities of a single observation.
Experiments report stronger cross-modal alignment and substantially greater sample efficiency than next-token prediction in the tested settings.
The results concern cross-modal alignment under specific assumptions and benchmarks, making their scope central to interpreting the broader claims.
Part of Marginalia: books, papers, and ideas for learning and debate.